[Train] Throw error messages if dataset_iter_config is not set in LightningTrainer.

[woshiyyya]

Description

User reports that they found an obscure particularity with the lightning trainer in combination with ray datasets:
The datasets_iter_config argument will be None if unset, but this will lead to an obscure error:

TypeError: An invalid dataloader was returned from `RayDataModule.train_dataloader()`. Found <torch.utils.data.dataloader.DataLoader object at 0x7f2280356710>.

We should raise an error with clear message to ask the users to provide a dataset_iter_config.

Use case

No response

[matthewdeng]

Should we just use a default value if the user passes in None?

[woshiyyya]

Yeah, we can pass a empty dict, which would use the default values defined in dataset.iter_torch_batches().

But I think it's better to make it mandatory when the users provided ray datasets, we don't want to silently set the default batch size for the users.